Multi-Modal Data Analysis for Pneumonia Status Prediction Using Deep Learning (MDA-PSP).

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Pulmonology

Background:

• Assessing pneumonia patient recovery using vital signs and chest X-rays (CXR) is challenging.
• Existing methods for pneumonia identification lack prediction of recovery status post-treatment.

Purpose of the Study:

• To develop a multimodal data analysis for pneumonia status prediction (MDA-PSP) to forecast hospital discharge within 7 days.
• To predict patient recovery status (improvement or deterioration) for early discharge decisions.

Main Methods:

• Designed a deep learning system (MDA-PSP) using vital signs and CXR images from admission days 1-3.
• Employed SHAP (SHapley Additive exPlanation), adaptive imputation, and dense layers-batch normalization (BN) with class weights.
• Utilized feature extraction from CXR images and vital signs for classification.

Main Results:

• The deep learning model achieved an accuracy of 0.77 for vital signs and 0.92 for CXR.
• The combined model (vital signs and CXR) demonstrated a prediction accuracy of 0.75.
• The MDA-PSP hybrid method showed higher prediction accuracy for pneumonia patient status.

Conclusions:

• Hybrid machine and deep learning methods offer a promising approach for predicting pneumonia patient discharge.
• The MDA-PSP system provides accurate predictions for patient recovery and discharge planning.
• Early prediction of pneumonia recovery status can optimize hospital resource allocation.